eval-under-loop: options for image allocation, direct I/O, mkfs and mount

[yarikoptic-gitmate]

Makes the loop backend usable for large (up to TB) images, and adds hooks for tuning the filesystem under test. Today the image is always dd'd from /dev/zero, which means writing every byte up front.

New options (each with an EVAL_UNDER_LOOP_* env var)

Option	What it does
--alloc zero|sparse|fallocate	zero is the default and keeps the old dd behaviour, so existing CI cells are unchanged. sparse uses truncate: instant, allocates nothing. fallocate calls fallocate(2) directly: it reserves the blocks without writing them. Where the host can't do it (NFS < 4.2, ramfs) it fails, instead of glibc's posix_fallocate quietly writing every block.
--direct-io	Attaches with losetup --direct-io=on, so image I/O isn't cached twice. If the host fs lacks O_DIRECT, current kernels refuse to attach; the backend now says that is why. Older kernels attach buffered instead, and the backend warns.
--mkfs-opts "..."	Extra mkfs.<fs> arguments, split on whitespace (e.g. btrfs -m single -K). There is no quoting.
--mount-opts a,b	Extra mount -o options. For vfat-like filesystems they are appended after uid=/gid=.

Also:

• The log shows the image's apparent vs allocated size, so you can tell a sparse image from a full one.
• --size is now checked to be a positive integer.
• With ro, the post-mount chown no longer aborts the run; it warns instead.

Docs: GOTCHAS.md's loop settings table now lists the allocation and loop-I/O settings, plus the btrfs minimum size of 109 MiB. The README has a large sparse btrfs example.

Behaviour for existing callers

With no new options, the run does the same thing as before. mount -o defaults is equivalent to a bare mount; a real ext4 run still gives rw,relatime. The only new output is one I: image on host: … log line. bin/ci/run-under.sh passes none of the new options.

Review and testing

• master is merged in, so the branch is up to date with master's NFS --mount-opts/--export-opts. That merge had no conflicts.
• Two independent reviewers, one test-driving and one reading the code, went through the first version. Their findings are fixed in the second commit:
  • the direct-I/O and fallocate claims in the docs were wrong (tmpfs has had O_DIRECT since Linux 6.6, and vfat implements fallocate by zeroing in the kernel);
  • ro mounts broke the post-mount chown;
  • newlines in --mkfs-opts silently dropped the rest of the string;
  • --size wasn't validated;
  • some help text still said the image is always dd'd;
  • the new options had no parse tests.
• Run for real as root on an ext4 host:
  • all three --alloc modes;
  • --direct-io (DIO=1), and on a ramfs host (clear error, clean teardown);
  • --mkfs-opts and --mount-opts, including invalid values, ro, and an embedded newline.
  • Teardown left no loop devices, mounts or image files behind in every case, failures included.
• bin/ci/shellcheck.sh is clean. bats tests/ passes 34/34, including 4 new unprivileged loop tests.
• Not tested: btrfs, vfat, xfs and NFS could not be mounted in the test container (its kernel lacks them). mkfs.btrfs itself was run: it rejects 100 MB (minimum is 114294784 bytes) and accepts 120 MB.

Not in this PR (follow-ups)

• A Loop btrfs matrix row. The git-annex target's loop-size-mb: 100 is below btrfs's minimum and would need raising.
• Nesting backends, e.g. eval-under nfs --no-root-squash -- eval-under loop --fs btrfs … to put the image on NFS. This already composes through TMPDIR, but it needs --no-root-squash, and turning it into a matrix row would need an over: field in evals/matrix.yaml.

🤖 Generated with Claude Code

https://claude.ai/code/session_0135xivQQnx5bWMzWD91J5TR

---

Generated by Claude Code